Please ignore until reviewed by @ChrisRackauckas. Opened as a draft.

Fixes #595.

Problem

sum(::AbstractVectorOfArray) was defined as:

function Base.sum(VA::AbstractVectorOfArray{T}) where {T}
    return sum(sum, VA.u)::T
end

The ::T assertion forces the result to equal the recursive element type T. That is wrong whenever summation does not return the element type:

• The reported ReverseDiff failure. For VectorOfArray([tunables]) under ReverseDiff.gradient, T = TrackedReal{Float64,Float64,<:TrackedArray} — each element carries its origin TrackedArray in the third type parameter. A computed sum, however, is a fresh TrackedReal{Float64,Float64,Nothing} (no origin), so the assertion throws:

  TypeError: in typeassert, expected ReverseDiff.TrackedReal{Float64, Float64, ReverseDiff.TrackedArray{…}},
  got a value of type ReverseDiff.TrackedReal{Float64, Float64, Nothing}
  

• It is not AD-specific. sum(VectorOfArray([Int8[1,2,3]])) also throws TypeError: expected Int8, got a value of type Int64 — ordinary integer promotion (Int8 → Int64) trips the same assertion.

Why the assertion existed

It was added in 69798c3 to stop Julia 1.10 LTS inference from widening the self-recursive sum(::AbstractVectorOfArray) (a VectorOfArray may contain VectorOfArrays) to Any, which is what @inferred sum(VA[VA[zeros(4,4)]]) guards against. Simply deleting ::T restores correctness but reintroduces that inference regression (verified: the nested case infers Any again).

Fix

Assert the type sum actually produces — the Base.add_sum accumulator type — instead of the element type:

function Base.sum(VA::AbstractVectorOfArray{T}) where {T}
    return sum(sum, VA.u)::Base.promote_op(Base.add_sum, T, T)
end

add_sum is the exact reduction operator Base.sum uses, so promote_op(add_sum, T, T) is Int64 for Int8, Float64 for Float64, and TrackedReal{…,Nothing} for ReverseDiff. promote_op is computed by inference over the scalar eltype (no VoA recursion), so it still pins the return type and keeps @inferred green on the LTS, while being correct for promoting and AD eltypes. If inference cannot determine add_sum(::T,::T) it returns Any, degrading gracefully to a no-op assertion (never wrong).

Tests

• test/interface_tests.jl: Int8/Float32 promotion (sum(VA[Int8[1,2,3]]) === Int64(6), …) plus @inferred.
• test/adjoints.jl: the ReverseDiff MWE from the issue (single- and nested-VoA gradients). ReverseDiff added to the test target (it is already a [weakdeps]/compat entry).

Verified locally on Julia 1.10.11, 1.11.9, 1.12.6, and 1.13.0-rc1 — all correctness and inference checks pass. Real test files pass: Partitions 115/115, Interface 139/139, Adjoint 13/13. Runic check is clean.

🤖 Generated with Claude Code